An exact algorithm for max-cut in sparse graphs

Transcription

1 An exact algorithm for max-cut in sparse graphs F. Della Croce a M. J. Kaminski b, V. Th. Paschos c a D.A.I., Politecnico di Torino, Italy b RUTCOR, Rutgers University c LAMSADE, CNRS UMR 7024 and Université Paris-Dauphine, France April 10, 2006 Abstract We study exact algorithms for the max-cut problem. Introducing a new technique, we present an algorithmic scheme that computes a maximum cut in graphs with bounded maximum degree. Our algorithm runs in time O (2 (1 (2/ ))n ). We also describe a max-cut algorithm for general graphs. Its time complexity is O (2 mn/(m+n) ). Both algorithms use polynomial space. Corresponding author. RUTCOR, Rutgers University, 640 Bartholomew Road, Piscataway, NJ 08854, USA, address: mkaminski@rutcor.rutgers.edu (Marcin Kaminski) 1

2 1 Background There has recently been a growing interest in analysis of the worstcase complexity of many NP-hard problems. Unless P NP, solving such problems requires super-polynomial time. Each problem in NP can be solved by a naive algorithm that exhaustively searches the solution space. However, for most of the problems more refined algorithms with better, but still exponential-time complexity, are known. Development of exact algorithms is mainly of theoretical interest but existence of fast exponential procedures may also have practical importance. Today s computers are able to handle moderate size instances of NP-hard problems. However, even though one can afford to run an exponential-time algorithm, polynomial-space complexity is a must. Satisfiability, graph coloring and maximum independent set are among the problems that have received much attention in the context of exact algorithms. In this paper we study another well-known NP-hard problem. Given an arbitrary graph with nonnegative weights assigned to its edges, the max-cut problem asks to find a partition of vertices into two subsets such that the sum of the weights of all the edges that have endpoints in two different parts of the partition is maximized. In the unweighted case (i.e. all weights are positive and equal) the problem is often referred to as simple 2

3 max-cut. Previous work. simple max-cut was one of the first problems whose NP-hardness was established. However, there are classes of graphs such as planar graphs, graphs with large girth ([6]), cographs and graphs with bounded treewidth ([1]), that admit polynomial-time solution of this problem. On the other hand, simple max-cut (and therefore max-cut) remains NP-hard even if restricted to such classes as chordal, split, or 3-colorable graphs ([1]). As shown in [13], the problem is NP-hard also for the class of graphs with bounded maximum degree, if 3. (For 2 the problem becomes trivial.) The worst-case complexity of the maximum cut problem has been studied in a few papers, some of them dealing with the weighted and some with the unweighted case. The fastest algorithm for simple max-cut in arbitrary graphs was proposed by Williams in [11]. In fact, the algorithm computes the number of solutions to an instance of the 2-CSP problem and employs interesting, non-standard techniques. (2-CSP is the constraint satisfaction problem with two variables per clause.) Used as a simple max-cut solver, the algorithm runs in time O (2 ωn/3 ) but, unfortunately, requires exponential space of O (2 ωn/3 ), where n is the number of vertices of the input 3

4 graph, ω < is the matrix multiplication exponent (the product of two k k matrices can be computed in time O(k ω )) and the notation O ( ) omits polynomial multiplicative terms. Whether there exists a polynomialspace algorithm that computes simple max-cut and runs faster than the naive one of time complexity O (2 n ) is an open question listed in [12]. More algorithms have been developed for sparse graphs. The upper bounds on their running times are given as linear functions of the number of edges in the input graph. (The number of edges of the input graph is denoted by m.) It makes them faster than the algorithms whose running time is bounded by a linear function of the number of vertices (like [11] or the naive algorithm) only if m is linearly bounded by n. In [3] an algorithm solving simple max-cut (via MAX-2-SAT) in time O (2 m/3 ) was proposed by Gramm et al. The bound was then improved to O (2 m/4 ) by Fedin and Kulikov in [2]. Their algorithm solves the maximum cut problem in a graph with integer weights on its edges. In a paper by Scott and Sorkin ([9]; see also [10]) a faster algorithm for max-cut, running in time O (2 min((m n)/2,m/5) ), was described. A recent paper by Kneis and Rossmanith ([4]) offers a simple max-cut algorithm with running time O (2 m/5.217 ). All of those algorithms use polynomial space. Our contribution. In this paper we develop a technique that seems to 4

5 be a new approach to the max-cut problem. The method consists in enumerating cuts in a subgraph H of G and then extending them in an optimal way to cuts in G. The technique is applied to graphs with bounded maximum degree and to general graphs. In both cases, we obtain an exponentialtime algorithm that uses polynomial space. For some classes of graphs our algorithms offer the best running time known. In particular, we obtain the fastest known algorithm solving the max-cut problem in the class of graphs with bounded maximum degree, if = 8, 9. We also provide a polynomial-space max-cut algorithm and a polynomial-space simple max-cut algorithm, that are the fastest known in the class of graphs with bounded maximum degree, for 8. For weighted graphs with bounded maximum degree, we present an algorithmic scheme that computes a maximum cut. For fixed, the algorithm runs in time O (2 (1 (2/ ))n ) and polynomial space. For 8, our algorithm is faster than the max-cut algorithm from [10] and the simple max-cut algorithm from [4]. It is slower than the exponential-space simple max-cut algorithm from [11] for 10. For general weighted graphs, we obtain an algorithm that computes a maximum cut and runs in time 2 mn/(m+n). Our algorithm is faster than the max-cut algorithm from [10] for m > 4n and faster than the simple 5

6 max-cut algorithm from [4] for m > 4.217n. It is slower than the simple max-cut exponential-space algorithm in [11] for m > ωn/(3 ω) > 3.808n. The organization of the paper is as follows. The next section is a formal introduction and contains definitions used later. In Section 3 we study a modification of the max-cut problem and develop our technique which is applied in Section 4 to graphs with bounded maximum degree and to general graphs in Section 5. 2 Introduction We consider weighted, undirected, loopless graphs without multiple edges. In a graph G = (V, E, w), V is the vertex set of cardinality V = n, E is the edge set of cardinality E = m, and w : E R + {0} is a weight function that assigns a nonnegative number w ij to each edge ij of G. The number of edges incident to a vertex in a graph is called the degree of the vertex. The maximum degree of all the vertices of a graph is called the maximum degree of the graph and denoted by. The average degree of a graph is the sum of degrees of all vertices of the graph divided by the number of its vertices. The average degree is denoted by d; notice that d = 2m/n. Given a subset U of vertices of V, the subgraph induced by the vertices in U is denoted by G[U]. 6

7 A cut C = (V 0, V 1 ) in a graph is a partition of its vertex set V into two disjoint subsets V 0 and V 1. The weight w(c) of cut C is the sum of weights of all the edges that have their endpoints in two different parts of the cut. Notice that the characteristic vector of one of the parts, say V 0, uniquely determines the partition. For the purpose of this paper, we will think of a partition as an assignment of 0 1 values to the vertices of the graph. Let x i be a Boolean variable which takes value 0, if v i V 0, and 1, if v i V 1. The weight of a cut in a graph G = (V, E, w) can be expressed as a pseudo-boolean function, (1) w(c) = w ij (x i x j + x i x j ) = w i x i 2 w ij x i x j, ij E i V ij E where w i = {i,j} E w ij. A maximum cut in a graph G is a cut of maximum weight. Given a graph G as an input, the max-cut problem asks to compute a cut in G that maximizes (1). Notice that it is enough to consider only connected graphs since if the graph is not connected, the max-cut problem can be solved for each of its connected components separately. It is easy to see that if the weights are restricted to be nonnegative real numbers, the max-cut problem can be solved in polynomial time for the class of bipartite graphs. 7

8 3 Extending a partial partition of vertices In this section we consider a modification of the max-cut problem. Suppose some of the vertices have already been partitioned into two subsets and now the problem is to find an optimal cut in the graph with respect to that pre-partition. We prove that if the graph induced by the vertices that have not yet been partitioned is bipartite, then the problem of finding an optimal extension of the partial partition can be solved in polynomial time. The algorithms presented in the following sections are based on this result. Let U V be a subset of vertices of G such that the subgraph G = G[U ] induced by the vertices in U = V \ U is bipartite. Also, let (U 0, U 1 ) be a partition of U into two subsets. Consider the problem of finding a partition (V 0, V 1 ) of V with U 0 V 0 and U 1 V 1 that maximizes (1). The vertices in U have already been assigned to some parts of the cut, thus variables x i, for i U, have their values fixed. There are four possible types of edges in the cut: edges with both endpoints in U, from U 0 to U, from U 1 to U, and with both endpoints in U. The problem of finding an optimal extension of the pre-partition is now equivalent to maximizing the following pseudo-boolean function, (2) w ij + w ijx j + w ijx j + w ij (x i x j + x i x j ) i U 0 i U 0 i U 1 i U j U 1 j U j U j U 8

9 where all sums are taken over edges ij E of the graph G. Putting, c j = i U 0 w ij i U 1 w ij + i U w ij where all sums are again taken over edges ij E, and omitting the constant term, the problem is equivalent to finding a maximum of the function, (3) c j x j 2 w ij x i x j j U ij E where E is the edge set of the bipartite graph G. In other words, the problem of finding an optimal extension of the pre-partition can be stated as the following integer quadratic program: (4) max c j x j 2 w ij x i x j j U ij E s.t. x i {0, 1} The standard linearization technique applied to (4) by introducing y ij = x i x j, yields the following integer linear program: (5) max c j x j 2 w ij y ij j U ij E s.t. y ij x i + x j 1 x i {0, 1} y ij {0, 1} It is easy to see that (4) and (5) are equivalent. They have the same optimal value and there is an easy correspondence between their optimal solutions, namely y ij = x i x j. 9

10 Having modelled the original quadratic problem (4) as an integer linear program, let us study the continuous relaxation of (5): (6) max c j x j 2 w ij y ij j U ij E s.t. y ij x i + x j 1 x i 0 x j 1 y ij 0 y ij 1 Lemma 1. The constraint matrix of the linear program (6) is totally unimodular, i.e., the determinant of every square submatrix of it equals 0 or ±1. Proof. Let A be the constraint matrix of (6). It has U + E columns and 2 U + 3 E rows and all its entries are either 0 or ±1. Let B be an edge-vertex incidence matrix of G, with rows corresponding to edges and columns corresponding to vertices. Notice that B is a submatrix of A. Moreover, any submatrix of A that has two non-zero entries in every row and every column has to be a submatrix of B. Take any square k k submatrix of A. We will prove the lemma by induction on k. Clearly, the result holds for k = 1. Now assume that all (k 1) (k 1) submatrices of A are totally unimodular and consider a matrix M which is a k k submatrix of A. 10

11 If all entries of any row or column of M are 0, then det(m) = 0 and M is totally unimodular. If any row or column of M has a single non-zero element (±1), then using the expansion method for calculating determinants and the induction hypothesis, it is easy to see that det(m) is either 0 or ±1, and A is totally unimodular. Suppose that each row and each column of M has at least two non-zero entries. Hence, M must be a submatrix of B but, since B is an incidence matrix of a bipartite graph, so is M. It is possible to partition the columns of M into two parts, according to the partition of vertices of bipartite graph. The sum of the columns in each part yields a unit vector (each edge of the bipartite subgraph has one endpoint in each part) and that implies linear dependence of M, therefore det(m) = 0 and M is totally unimodular. Theorem 2. Let U V be such that the subgraph G = G[U ] induced by the vertices in U = V \ U is bipartite and (U 0, U 1 ) be a partition of U into two subsets, then the problem of finding a partition (V 0, V 1 ) of V with U 0 V 0 and U 1 V 1 that maximizes (1) is polynomial-time solvable. Proof. The problem of finding a partition (V 0, V 1 ) of V with U 0 V 0 and U 1 V 1 that maximizes (1) can be modelled as the integer quadratic program (4) which is equivalent to (5). Total unimodularity of the constraint matrix of (6) (by Lemma 1) implies the existence of an optimal 0 1 solution 11

12 of (6), and such a solution can be found in polynomial time (see for example [8]). Since the relaxation (6) of (5) has an optimal 0 1 solution, therefore (4) can be solved in polynomial time. Before we proceed to the next section, let us briefly describe the algorithmic technique we are going to apply. Given an induced bipartite subgraph G[B] of G, one can enumerate all partitions of V \ B and find an optimal extension of each in polynomial time (by Theorem 2). The complexity of such a technique is O (2 V \B ) and it strongly depends on the size of the bipartite subgraph that has to be constructed. 4 Algorithm for graphs with bounded maximum degree In this section we present and analyze an algorithmic scheme A( ). For a fixed integer ( 3), the scheme yields an algorithm whose input is a weighted graph G = (V, E, w) of maximum degree and whose output is a maximum cut in G with respect to the weight function w. Step 1. If G is isomorphic to the complete graph on + 1 vertices, then let B be any pair of vertices and go to Step 3. 12

13 Step 2. -color G. Let B be the union of 2 largest color classes of the coloring. Step 3. Enumerate all partitions of elements of V \ B into two subsets (all 0 1 assignments) and for each find an optimal extension of the partial partition. Step 4. Find a cut C that has the largest weight among all checked in Step 3. Return the cut C. Theorem 3. For a fixed integer ( 3), Algorithm A( ) computes maxcut in a graph G in time O (2 (1 (2/ ))n ) and polynomial space. Proof. Let us first notice that the algorithm indeed finds a maximum cut. It is clear that the induced subgraph G[B] is bipartite. Therefore, any partition of V \ B into two subsets can be extended to an optimal partition of V in polynomial time by Theorem 2. Clearly, by enumerating all partitions of V \ B and then extending each in an optimal way, one finds a maximum cut in G. The enumeration of partitions in Step 3 is the bottleneck of the algorithm; it needs exponential time O (2 V \B ). Other steps can be performed in linear time. It is clear for Steps 1 and 4, and the linear time algorithm for Step 2 is given in [5]. Notice, that the algorithm can be implemented 13

14 in such a way that each step uses only polynomial space. In particular, in Step 3 we need to store only the current best solution. Suppose that the input graph is isomorphic to the complete graph on + 1 vertices. The number of partitions that are enumerated in Step 3 is 2 n 2 but since = O(n) the claimed running time follows. Now suppose that the input graph G is not isomorphic to the complete graph on + 1 vertices. Then, by Brooks Theorem G is colorable ([5]). Clearly, the union of two largest color classes has size at least 2n/ and V \ B n(1 (2/ )). The number of partitions that are enumerated in Step 3 is O (2 (1 (2/ ))n ) and the claimed running time follows. 5 Algorithm for general graphs Let us notice that in the algorithm presented in the previous section, the assumption of bounded maximum degree is needed only to obtain an induced bipartite graph. Now we relax this assumption and study the complexity of the method in general graphs. Let us formalize that as Algorithm B. The input of B is a weighted graph G = (V, E, w) and the output is a maximum cut in G with respect to the weight function w. Step 1. Find a maximal independent set I 0 in G. 14

15 Step 2. Find a maximal independent set I 1 in G[V \ I 0 ]. Let B be the union of I 0 and I 1. Step 3. Enumerate all partitions of elements of V \ B into two subsets (all 0 1 assignments) and for each find an optimal extension of the partial partition. Step 4. Find a cut C that has the largest weight among all checked in Step 3. Return the cut C. To complete the description of the algorithm, we need to provide a procedure that finds an induced bipartite subgraph in Steps 1 and 2. From Turan s theorem follows that the size of a maximum independent set is at least n/(d + 1), and as shown in [7], there is a linear-time algorithm that constructs an independent set of at least that size. As the time complexity of B depends on B, we need to give a lower bound on the size of the bipartite subgraph B. Claim 4. The set B of vertices constructed in Step 2 of Algorithm B has size at least 2/(d + 2). Proof. Let i = I 0 and m be the number of edges of the subgraph G[I 0 I 1 ]. If i 2n/(d + 2), then B 2n/(d + 2) and the claim follows. Suppose i < 2n/(d + 2). The average degree d in the graph G[V I 0 ] 15

16 is d = 2(m m )/(n i). Notice that m n i, since I 0 is an independent set. Hence, d 2n/(n i) 2 and since i < 2n/(d + 2), we have d < d. It follows that B = i + (n i)/(d + 1) 2n/(d + 2). Having established the lower-bound on the size of B, we can claim the running time of Algorithm B. Notice that 2/(d+2) = n/(n+m) and n B mn/(m + n). Theorem 5. Algorithm B computes max-cut in a graph G with n vertices and m edges in time O (2 mn/(m+n) ), and polynomial space. The proof of this theorem is similar to the proof of Theorem 3 and will be omitted. 6 Acknowledgments We thank the anonymous referees for their helpful comments on the paper. The work was done while the first author was visiting LAMSADE on a research position funded by CNRS and the second author was visiting LAMSADE being supported by DIMACS under the NSF grant INT to Rutgers University. The second author would like to thank Prof. Michael Saks for introduc- 16

17 ing him to the field of exact algorithms. His thanks also go to LAMSADE for their hospitality and DIMACS for the support. References [1] H. L. Bodlaender and K. Jansen. On the complexity of the maximum cut problem. Nordic Journal of Computing, 7(1):14 31, [2] S. S. Fedin and A. S. Kulikov. A 2 E /4 -time algorithm for max-cut. Journal of Mathematical Sciences. To appear. Available at logic.pdmi.ras.ru/~kulikov/maxcut_e.ps.gz. [3] J. Gramm, E. A. Hirsch, R. Niedermaier, and P. Rossmanith. Worstcase upper bounds for Max2Sat with an application to MaxCut. Discrete Appl. Math., 130: , [4] J. Kneis and P. Rossmanith. A new satisfiablity algorithm with applications to max-cut. Technical Report AIB , Department of Computer Science, RWTH Aachen [5] L. Lovász. Three short proofs in graph theory. J. Combin. Theory Ser. B, 19: ,

18 [6] S. Poljak and Z. Tuza. Maximum cuts and large bipartite subgraphs. In W. Cook, L. Lovász, and P. Seymour, editors, Combinatorial Optimization. Papers from the DIMACS Special Year, pages AMS, [7] S. Sakai, M. Togasaki, and K. Yamazaki. A note on greedy algorithms for the maximum weighted independent set problem. Discrete Appl. Math., 126: , [8] A. Schrijver Theory of linear and integer programming. Wiley, [9] A. D. Scott and G. B. Sorkin. Faster algorithms for MAX CUT and MAX CSP, with polynomial expected time for sparse instances. In Proc. RANDOM 03, volume 2764 of Lecture Notes in Computer Science, pages Springer-Verlag, [10] A. D. Scott and G. B. Sorkin. Solving sparse semi-random instances of Max-Cut and Max-CSP in linear expected time. Research Report (W ), IBM Research division, Thomas J. Watson Researcc Center, [11] R. Williams. A new algorithm for optimal 2-constraint satisfaction and its implications. Theoretical Computer Science, Vol. 348, 2-3, p

19 [12] G. J. Woeginger. Space and time complexity of exact algorithms: some open problems. In Proc. IWPEC 04, volume 3162 of Lecture Notes in Computer Science, pages Springer-Verlag, [13] M. Yannakakis. Node- and edge-deletion NP-complete problems. In Proc. STOC 78, pages ,